Tiu, Ervin Shan Khai (2019) Data Pre-Processing Techniques For Improving River Water Level Prediction: A Case Study Of The Dungun River, Terengganu, Malaysia. Master dissertation/thesis, UTAR.
Abstract
An accurate river water level prediction model is vital for the development of flood mitigation plans in a river basin, and the accuracy of input data is important for ensuring good predictions. In this research study, a Support Vector Regression (SVR) model was applied to predict river water levels at the Dungun River, Terengganu, Malaysia. A major challenge of this study is to process the observed rainfall series which may tend to be incomplete and inconsistent. A better observed rainfall series data can lead to improved performance of the river water level prediction model. This study adopted the data pre-processing techniques to first improve the observed rainfall series. Three data pre-processing techniques namely; the Variational Mode Decomposition (VMD) method, the Boosting method and the Boosting-VMD method, were adopted to pre-process the observed rainfall series at the basin prior to applying them into the prediction model. Rainfall series and river water levels from November to February (Northeast Monsoon period) for the period of 1996-2016 were used. The three pre-processed rainfall series and the original observed rainfall series were then separately and individually, used to predict river water levels using the SVR model. The predicted river water levels from the four modified SVR models namely; the Ori-SVR (using original observed rainfall) model, the VMD-SVR (using VMD method) model, the B-SVR (using Boosting method) model, and the B-VMD-SVR (using Boosting-VMD method) model, were assessed against the observed river water levels, using non-parametric statistics (Model prediction errors’ range, standard deviation and confidence interval range) and parametric statistics (Bias, Root-Mean-Square Error, Mean Absolute Percentage Error, Nash-Sutcliffe Efficiency Coefficient and Mean Absolute Error) where appropriate, to assess the three data pre-processing techniques in enhancing the SVR model. Results indicated that the VMD, the Boosting and the Boosting-VMD methods afforded data improvements that boosted the performance of SVR models. Further statistical analyses showed that the SVR model cum the Boosting-VMD method that is, the (B-VMD-SVR) to be the most robust, with results of model prediction error’s range (4.27; Min: -2.95, Max: 1.32), model prediction error’s standard deviation (0.4200), model prediction error’s confidence interval range ([-0.02, 0.01]; Standard error: 0.00868), Bias (0.00), Root-Mean-Square Error (0.42), Mean Absolute Percentage Error (4.36), Nash-Sutcliffe Efficiency Coefficient (0.96) and Mean Absolute Error (0.28).
Actions (login required)
![[img]](http://eprints.utar.edu.my/style/images/fileicons/application_pdf.png)
